Wuying-Browser-Agent Framework Bridges Real-World Browser Automation Gap
The Wuying-Browser-Agent framework has been designed for browser agents that operate on live websites over extended periods. Its developers stress that scaling alone isn't sufficient; it's crucial to ensure alignment at every phase of the process—like execution, supervision, optimization, and evaluation. The framework includes a structured browser harness that provides reliable execution and good context management. For training on recovery paths and complex UI interactions, they've introduced the Reflection and UI-specialized Curriculum SFT (RUIC-SFT). There's also the Divergence-Aware Online GRPO (DAO-GRPO), which enhances long-term credit assignment using potential-based reward shaping and step weighting. To track progress, the team has created BrowserBench, a bilingual benchmark with 350 tasks to demonstrate the challenges faced in real-world applications.
Key facts
- Wuying-Browser-Agent is a unified framework for real-world browser automation.
- The framework addresses execution, supervision, optimization, and evaluation alignment.
- A structured browser harness provides stable execution primitives and decision-oriented context management.
- RUIC-SFT trains explicitly on recovery trajectories and complex-UI interactions.
- DAO-GRPO improves long-horizon credit assignment through potential-based reward shaping and divergence-aware step weighting.
- BrowserBench is a bilingual real-web benchmark of 350 tasks.
- Browser agents perform well on short, clean demonstrations but not in real deployment.
- Real deployment requires dozens of decisions, mistake recovery, and complex UI navigation.
Entities
—